Improving the detection accuracy of linear clear-cuts for effective satellite monitoring of illegal forest felling

This work is devoted to the problem of automated detection of forest fellings based on Earth remote sensing data in the context of timely measures to protect forest legislation. Manual image analysis is inefficient due to human fatigue caused by prolonged monotonous work, leading to increased interest in neural network-based monitoring systems enabled by recent advances in artificial intelligence technology. Such solutions combining human perception capabilities with the potential of neural networks provide high accuracy and speed in processing large volumes of satellite data, thus improving the effectiveness of measures aimed at protecting forest resources. Although neural network methods have been successfully applied in the considered field, their practical implementation encounters substantial challenges stemming from insufficient performance in detecting illegally cleared areas designated for constructing roads, power lines, and pipelines. This makes it impossible to detect a significant portion of forestry law violations. Aim of the study. Improving the accuracy of linear forest clear-cuts recognition in satellite images without reducing the quality of recognition of other types of felling. Materials and methods. A method is proposed based on neural network ensemble learning, which incorporates a modified Tversky Loss function for model training and applies bitwise aggregation of outputs. Its efficacy was tested through experiments on a dataset of forest logging activities in Khanty-Mansi Autonomous Okrug – Yugra between 2018 and 2022. This dataset includes labeled Sentinel-2 satellite imagery covering both “snowless” (June–September) and “snowy” (November–April) felling seasons. Results. Proposed method improved linear forest clear-cuts detection accuracy by 5.35 % for “snowless” season and by 6.8 % for “snowy” season, with no decrease in recognition quality for other clearing types. Conclusion. Obtained results provide a foundation for future research targeting hard-to-detect felling activities, especially those concealed by dense clouds, haze, or cloud shadows.